Of course, the speed estimate could be wrong if the bandwidth bottleneck (on either previous sites or the current one) is further than the last mile, especially for e.g. Chinese users whose Internet speeds are drastically slower with international peers than domestic ones. Or alternately if the user is on a mobile connection and is moving quickly between areas with good and poor reception, going on and off Wi-Fi, etc., although in that case the browser could incorporate signal strength data from the OS. Still, on an Internet whose users have wildly varying connection speeds, I think it makes sense to allow leveraging the difference somehow.